OpenAI next flagship uses looped transformer, report says

Looped transformers could lift model capability while complicating alignment and safety oversight.

A report says OpenAI's next flagship uses the same looped transformer technique as Astra to boost math and coding, but looping makes safety monitoring harder; Microsoft leaked details on a public page.

Read it at The Decoder →

Coverage: 2 sources, in order

1. 14 Oct 16:14 The Decoder: OpenAI will reportedly ship at least one model this year but its flagship still needs alignment work
2. 14 Oct 16:18 The Decoder: OpenAI will reportedly ship at least one model this year but its flagship still needs alignment work

← Front page